Using Prior Knowledge to Calibrate a Radio Frequency Frontend

Abstract

A method can comprise, as part of initially deploying a second radio frequency (RF) transceiver, transferring, by a system, a core model to the second RF transceiver, the core model having been trained based on a training process comprising training an artificial intelligence model for a first RF transceiver, based on first data that is measured for the first RF transceiver. The method can further comprise applying, by the system, transfer learning on the core model at the second RF transceiver based on second data that is measured for the second RF transceiver, to produce a trained model. The method can further comprise calibrating, by the system, the second RF transceiver based on an output of the trained model to produce a calibrated second RF transceiver. The method can further comprise transmitting, by the system, RF information via the calibrated second RF transceiver.

What is claimed is: 
     
         1 . A method, comprising:
 as part of initially deploying a second radio frequency transceiver, transferring, by a system, a core model to the second radio frequency transceiver, the core model having been trained based on a training process comprising training an artificial intelligence model for a first radio frequency transceiver, based on first data that is measured for the first radio frequency transceiver, wherein the core model identifies a number of layers and respective numbers of neurons for respective layers of the number of layers, and wherein the core model is transferred independent of identifying interconnections and coefficients;   applying, by the system, transfer learning on the core model at the second radio frequency transceiver based on second data that is measured for the second radio frequency transceiver, to produce a trained model, wherein the trained model comprises the number of layers, the respective numbers of neurons for respective layers of the number of layers, the interconnections, and the coefficients;   calibrating, by the system, the second radio frequency transceiver based on an output of the trained model to produce a calibrated second radio frequency transceiver; and   transmitting, by the system, radio frequency information via the calibrated second radio frequency transceiver.   
     
     
         2 . The method of  claim 1 , wherein a group of radio frequency transceivers comprises the second radio frequency transceiver, wherein the output is a first output, and further comprising:
 transferring, by the system, respective copies of the core model to respective radio frequency transceivers of the group of radio frequency transceivers;   applying, by the system, respective transfer learning on the respective copies of the core model at the respective radio frequency transceivers based on respective data that is measured for the respective radio frequency transceivers, to produce respective trained models; and   calibrating, by the system, the respective radio frequency transceivers based on respective outputs of the respective trained model to produce respective calibrated radio frequency transceivers.   
     
     
         3 . The method of  claim 1 , wherein transferring the core model to the second radio frequency transceiver comprises:
 copying, by the system, learned weights of the core model and copying the core model from the first radio frequency transceiver to the second radio frequency transceiver.   
     
     
         4 . The method of  claim 1 , wherein transferring the core model to the second radio frequency transceiver is performed based on the second radio frequency transceiver being powered on before carrying live traffic. 
     
     
         5 . The method of  claim 1 , wherein training the artificial intelligence model for the first radio frequency transceiver was performed based on generating a group of data sets by varying radio frequency-platform specific parameters, while holding system-related parameters constant, and by varying values of the group of data sets within a target operational signal-to-noise ratio. 
     
     
         6 . The method of  claim 5 , wherein the radio frequency-platform specific parameters comprise frequency offset or quadrature-phase skew, and wherein the system-related parameters comprise signal bandwidth, modulation order, or calibrated power amplifier bias. 
     
     
         7 . The method of  claim 1 , wherein the artificial intelligence model is trained using fully-supervised learning with first labeled data, and transfer learning on the core model is performed using partially-supervised learning with second labeled data and part of the first data. 
     
     
         8 . The method of  claim 1 , wherein the transfer learning is applied on the core model to update coefficients of the trained model while holding a structure of the core model constant in the trained model. 
     
     
         9 . A system, comprising:
 a processor; and   a memory coupled to the processor, comprising instructions that, in response to execution by the processor, cause the system to perform operations, comprising:
 as part of initially deploying a second transceiver, transferring a core model to the second transceiver, the core model having been trained, based on first data that is measured for a first transceiver, using an artificial intelligence model for the first transceiver, wherein the core model identifies a number of layers and respective numbers of neurons for respective layers of the number of layers; 
 performing transfer learning on the core model at the second transceiver based on second data that is measured for the second transceiver, to produce a trained model; 
 calibrating the second transceiver based on an output of the trained model to produce a calibrated second transceiver; and 
 transmitting information via the calibrated second transceiver. 
   
     
     
         10 . The system of  claim 9 , wherein using the artificial intelligence model comprises using first labeled training data, wherein performing the transfer learning using the core model comprises performing the transfer learning with second labeled training data, and wherein the first labeled training data differs from the second labeled training data. 
     
     
         11 . The system of  claim 9 , wherein a first behavior of the first transceiver differs from a second behavior of the second transceiver based on differing hardware effects and channel effects. 
     
     
         12 . The system of  claim 9 , wherein a first behavior of the first transceiver differs from a second behavior of the second transceiver based on a first channel condition of the first transceiver differing from a second channel condition of the second transceiver, or based on a first hardware configuration of the first transceiver differing from a second hardware configuration of the second transceiver. 
     
     
         13 . The system of  claim 9 , wherein performing the transfer learning on the core model at the second transceiver based on second data that is measured for the second transceiver, to produce the trained model comprises:
 updating coefficients of the core model in the trained model while holding a structure of the core model constant in the trained model.   
     
     
         14 . The system of  claim 9 , wherein calibrating the second transceiver comprises:
 adjusting a transmit output power, an energy efficiency, a skew, or a frequency offset of the second transceiver.   
     
     
         15 . A non-transitory computer-readable medium comprising instructions that, in response to execution, cause a system comprising a processor to perform operations, comprising:
 training a model for a first transceiver, based on first data that is measured for the first transceiver, to produce a core model;   transferring the core model to a second transceiver;   performing transfer learning on the core model at the second transceiver based on second data that is measured for the second transceiver, to produce a trained model; and   calibrating the second transceiver based on an output of the trained model to produce a calibrated second transceiver.   
     
     
         16 . The non-transitory computer-readable medium of  claim 15 , wherein the operations further comprise:
 transmitting information via the calibrated second transceiver.   
     
     
         17 . The non-transitory computer-readable medium of  claim 15 , wherein calibrating the second transceiver comprises:
 adjusting at least one of a transmit output power of the second transceiver, an energy efficiency of the second transceiver, a skew of the second transceiver, or a frequency offset of the second transceiver.   
     
     
         18 . The non-transitory computer-readable medium of  claim 15 , wherein performing transfer learning on the core model at the second transceiver comprises:
 modifying coefficients in the trained model while holding constant a structure of the trained model.   
     
     
         19 . The non-transitory computer-readable medium of  claim 15 , wherein performing the transfer learning on the core model at the second transceiver based on second data that is measured for the second transceiver, to produce the trained model is performed based on a quantized neural network architecture that utilizes L-bit representations for data and Q-bit representations for weight. 
     
     
         20 . The non-transitory computer-readable medium of  claim 15 , wherein the output of the trained model is a first output of the trained model, and wherein the operations further comprise:
 updating the trained model based on a physical change to a radio that comprises the second transceiver, to produce an updated trained model; and   re-calibrating the second transceiver based on a second output of the trained model to produce an updated calibrated second transceiver.